The Open Cloud Testbed (OCT) is an NSF-funded CIRC project whose goal is to provide bare-metal servers to the research community. In contrast to other testbeds that offer CPUs and GPUs, OCT offers a large set of FPGAs directly connected to the network that can be programmed by testbed users.

During the initial phase, OCT consisted of eight AMD Alveo U280 accelerator cards configured with direct network connectivity. Pre-compiled TCP and UDP network stack binaries for the FPGAs were provided to support experiments involving network-attached FPGAs. User support primarily focused on the AMD Vitis flow, which enabled users to build their projects on top of a pre-programmed FPGA shell provided by AMD. Users could gain access to these FPGAs by joining a designated CloudLab project, with FPGA resources provisioned through a CloudLab profile. Support for projects involving custom shells such as OpenNIC and Corundum remained restricted during this stage, and access for those users was granted through a separate project.

Subsequent expansions significantly increased the capabilities of OCT through the addition of sixteen more U280 accelerators, bringing the total to 24. Along with this expansion, a unified workflow was created to support both Vitis and custom flows, eliminating the need to maintain a separate project for users who use the custom flow. Supporting Vivado users through a separate project required manual administrative intervention to execute factory resets between experiments, which created significant management challenges. To resolve this, an automated workflow was implemented to factory reset the FPGAs independent of the workflow. Specifically, an FPGA is programmed with a golden shell at experiment termination. As a result, the next user always receives a clean bare-metal FPGA free of any prior shell configuration. This approach successfully supports both Vivado and Vitis flows within a single project framework with minimal manual intervention.

While the U280 has long served as the backbone of the FPGAs in OCT, our infrastructure has steadily evolved to support a diverse range of workloads including machine learning and AI. To address these emerging applications, OCT has incorporated AMD VCK5000 and V70 accelerators that are based on the Versal Adaptive SoC architecture. These devices come with specialized AI engines that are capable of accelerating modern workloads such as real-time object detection, image classification, and video analytics.